Modeling Cognition through Adaptive Neural Synchronization: A Multimodal Framework Using EEG, fMRI, and Reinforcement Learning

Rashad Hall¹Horace T Crogman^1*Maryam Maleki¹Maury Damon Jackson²

• ¹Department of Physics, California State University Dominguez Hills, Carson, CA 90747, Carson, United States
• ²H.M.S Richard Divinity School, La Sierra University, Riverside CA 92505, Riverside, United States

Understanding the cognitive process of thinking as a neural phenomenon remains a central challenge in neuroscience and computational modeling. This study presents a biologically grounded framework that integrates neuronal synchronization, metabolic energy cost, and reinforcement learning to simulate adaptive decision-making across cognitive states. Neural synchronization is modeled using Kuramoto oscillators driven by real EEG and fMRI signals, while energy dynamics are constrained by multimodal brain activity profiles. Reinforcement learning agents (Q-learning and DQN) learn to modulate external inputs to maintain optimal synchrony under energy-efficient conditions. The agents achieved rapid convergence, with DQN stabilizing cumulative rewards within 200 episodes and reducing mean synchronization error by over 40%, outperforming Q-learning in both speed and generalization. The model successfully reproduced canonical brain states—focused attention, multitasking, and rest—and was validated against empirical EEG-fMRI data. Spectral analysis revealed alpha-band power of 3.2×10⁻⁴ a.u. in simulated data versus beta-dominance (3.32×10⁻⁴ a.u.) in real EEG, indicating fidelity to resting states and tunability for task-related engagement. Phase Locking Value (PLV) analysis showed strong alpha synchrony (0.9806–0.9926) and lower circular variance in the focused condition (0.0456), with a near-significant phase shift compared to rest (t = –2.15, p = 0.075). Cross-modal validation showed moderate correlation between simulated and real BOLD signals (r = 0.30 in the resting condition), with delayed input alignment improving temporal dynamics. Simulated BOLD signals analyzed via General Linear Models (GLMs) achieved high region-specific prediction accuracy (R² = 0.973–0.993, p < 0.001), especially in prefrontal, parietal, and anterior cingulate cortices. Voxel-wise correlation heatmaps and ICA decomposition confirmed structured network dynamics, consistent with empirical brain systems. These results demonstrate that the model bridges electrophysiological timing with spatial network activation, captures neuroenergetic constraints, and adaptively regulates brain-like states through learning. This framework offers a scalable foundation for modeling cognition and designing biologically inspired neuroadaptive systems.